Imitation Learning for Path Planning in Cardiac Percutaneous Interventions.

Objective: Mitral regurgitation is a valvular heart disease particularly affecting the aging population. Minimally invasive transcatheter procedures offer benefits over traditional open-chest surgery but require significant operator skill and hand-eye coordination, making the learning curve steeper and limiting accessibility. To address these challenges, there is growing research interest in automating these procedures, making it crucial to define safe navigable routes within anatomical structures for robotic operation. This study introduces a tailored learning-based framework for path planning in cardiac percutaneous interventions, specifically adapted to the dynamically constrained and safety-critical environment of mitral valve repair.

Methods: We compared generative adversarial imitation learning and behavioral cloning techniques to traditional path planning algorithms like rapidly-exploring random trees. Using patient-specific anatomical data, a faithful digital twin was created, with dynamic motions to replicate real-time cardiac movements of the mitral valve.

Results: Learning approaches significantly reduced target position errors and improved path smoothness with greater clearance from obstacles compared to state-of-the-art methods.

Conclusion: Learning methodologies provided consistent and repeatable routes in cardiac anatomy, both in pre-operative static and intra-operative dynamic scenarios.

Significance: Embedding task demonstrations in the learning process shows the potential to automate and optimize catheter navigation, promoting standardization of minimally invasive cardiac procedures.